Emotional Resonance Matching for Personalized E-Commerce Re-Engagement System Using Machine Learning and Generative AI

Abstract

Cart abandonment remains one of the most persistent revenue leakages in e-commerce, with industry estimates placing the global abandonment rate above 70 percent. This paper describes the design and implementation of an end-to-end customer re-engagement system that draws on AWS cloud infrastructure, supervised machine learning, and large-language-model-based email generation. At its core, a gradient-boosted classifier trained on anonymised clickstream and purchase records predicts the discount tier most likely to convert each abandoned-cart session into a completed purchase. The predicted offer is then passed to a generative AI copywriter that drafts a personalised email whose tone adapts to the sentiment extracted from the shopper’s past product reviews—a capability we call Emotional Resonance Matching. An event-driven pipeline built on AWS Lambda, Step Functions, Amazon Pinpoint, Amazon S3, AWS Glue, and Amazon Redshift orchestrates the entire flow from trigger to delivery. Offline evaluation shows a 23 percent uplift in simulated click-through rate over a rule-based baseline, and the median end-to-end latency from cart-abandonment event to email delivery sits comfortably under 90 seconds.

Introduction

It begins by highlighting that around 70% of online shopping carts are abandoned, leading to major revenue loss. While current recovery methods rely on generic or template-based emails, they fail to fully utilize customer behavioral data or personalize communication effectively. The proposed system addresses this gap by combining machine learning and large language models to optimize both discount offers and message tone.

The system introduces a pipeline where a cart-abandonment event triggers an automated workflow. Customer data (such as RFM metrics, browsing behavior, and review sentiment) is processed using cloud services like AWS EventBridge, Step Functions, Redshift, and SageMaker. A machine learning model (XGBoost) predicts the most suitable discount tier (0%, 5%, 10%, or 15%), while an LLM generates personalized email content.

A key innovation is Emotional Resonance Matching (ERM), which adjusts the tone of the email based on the customer’s historical sentiment—making messages more formal or more casual depending on user behavior.

Prior research is cited showing that faster follow-up increases conversion rates and that ML-based discount optimization outperforms rule-based systems. The system is designed for low-latency response (under 90 seconds) and scalable cloud deployment using serverless architecture.

Conclusion

We presented an automated e-commerce customer re-engagement system that combines event-driven cloud infrastructure, supervised ML for discount-tier selection, and generative AI for personalised email copy. The system’s distinguishing feature, Emotional Resonance Matching, conditions the LLM copywriter’s tone on sentiment inferred from each customer’s review history, producing messages that feel qualitatively more aligned with how individual shoppers communicate. Offline evaluation confirms meaningful uplifts over rule-based baselines in both offer prediction accuracy and tone appropriateness. Future work will focus on three areas: replacing the nightly Glue batch with a streaming feature store to eliminate the latency bottleneck identified in Section VII-D; extending the ERM sentiment model to handle multilingual review text; and introducing an online reinforcement-learning layer that updates the offer policy in response to real conversion feedback, closing the loop that the current supervised model leaves open.

References

[1] Baymard Institute, “C art abandonment rate statistics,” 2023. [Online]. Available: https://baymard.com/lists/cart-abandonment-rate [2] P. A. Pavlou and M. Fygenson, “Understanding and predicting electronic commerce adoption,” MIS Quarterly, vol. 30, no. 1, pp. 115–143, 2006. [3] L. Xu, J. Duan, and A. Whinston, “Path to purchase: A mutually exciting point process model for online advertising and conversion,” Management Science, vol. 60, no. 6, pp. 1392–1412, 2014. [4] F. Provost and T. Fawcett, Data Science for Business. Sebastopol, CA: O’Reilly Media, 2013. [5] L. Prokhorenkova, G. Gusev, A. Vorobev, A. V. Dorogush, and A. Gulin, “CatBoost: unbiased boosting with categorical features,” in Advances in Neural Information Processing Systems, 2018, pp. 6638–6648. [6] V. Mnih, K. Kavukcuoglu, D. Silver et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, pp. 529–533, Feb. 2015. [7] T. B. Brown, B. Mann, N. Ryder et al., “Language models are few-shot learners,” in Advances in Neural Information Processing Systems, 2020, vol. 33. [8] L. Ouyang, J. Wu, X. Jiang et al., “Training language models to follow instructions with human feedback,” in Advances in Neural Information Processing Systems, 2022. [9] E. Jonas, J. Schleier-Smith, V. Sreekanti et al., “Cloud programming simplified: A Berkeley view on serverless computing,” arXiv:1902.03383, 2019. [10] Amazon Web Services, “AWS Step Functions Developer Guide,” 2024. [Online]. Available: https://docs.aws.amazon.com/step-functions/

Copyright

Copyright © 2026 Samundeeswari M, Logapriya B, Priyadarshini K, Nityasri R S . This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.